Update repository

Co-authored-by: laurier-rochon <laurier-rochon@users.noreply.huggingface.co>
Co-authored-by: pveugen <pveugen@users.noreply.huggingface.co>
Co-authored-by: akagabi <akagabi@users.noreply.huggingface.co>

.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,316 @@
+---
+license: other
+license_name: desert-ant-labs-source-available-1.0
+license_link: https://license.desertant.com/1.0
+language:
+- af
+- am
+- ar
+- as
+- az
+- be
+- bg
+- bn
+- bo
+- ca
+- chr
+- cs
+- cy
+- da
+- de
+- dv
+- el
+- en
+- eo
+- es
+- et
+- eu
+- fa
+- fi
+- fr
+- ga
+- gl
+- gu
+- he
+- hi
+- hr
+- hu
+- hy
+- id
+- is
+- it
+- ja
+- ka
+- kk
+- km
+- kn
+- ko
+- ku
+- ky
+- la
+- lo
+- lt
+- lv
+- mk
+- ml
+- mn
+- mr
+- ms
+- my
+- ne
+- nl
+- no
+- or
+- pa
+- pl
+- pt
+- ro
+- ru
+- si
+- sk
+- sl
+- sq
+- sr
+- st
+- sv
+- sw
+- syr
+- ta
+- te
+- th
+- tl
+- tr
+- ug
+- uk
+- ur
+- vi
+- xh
+- yo
+- zh
+tags:
+- language-identification
+- language-detection
+- langid
+- text-classification
+- on-device
+- edge-ai
+- mobile
+- browser
+- offline
+- onnx
+- multilingual
+pipeline_tag: text-classification
+---
+
+<!-- card-header:start (generated from manifest.json, edit below this block) -->
+# Tongue
+
+Detect language based on 3 words.
+
+On-device language identification for short text across 84 languages.
+
+- **SDKs, install and examples:** https://github.com/Desert-Ant-Labs/desert-ant-core/blob/main/docs/models/tongue.md
+- **Website:** https://desertant.com/models/tongue/
+
+<!-- card-header:end -->
+
+<!-- card-install:start (generated from manifest.json, edit below this block) -->
+| | |
+| --- | --- |
+| **Platforms** | iOS, macOS, tvOS, visionOS, Android, Linux, Windows, Browser, Node |
+| **Languages** | 84 |
+| **Weights** | [v1.0.0](https://huggingface.co/desert-ant-labs/tongue) |
+
+## Install
+
+**Swift** ([requirements](https://github.com/Desert-Ant-Labs/desert-ant-core/blob/main/README.md#swift))
+
+```swift
+.package(url: "https://github.com/Desert-Ant-Labs/desert-ant-core.git", from: "3.1.0")
+```
+
+Then add the `Tongue` product to your target.
+
+**Kotlin** ([requirements](https://github.com/Desert-Ant-Labs/desert-ant-core/blob/main/README.md#android))
+
+```kotlin
+implementation("ai.desertant:tongue:3.1.0")
+```
+
+**JavaScript** ([requirements](https://github.com/Desert-Ant-Labs/desert-ant-core/blob/main/README.md#javascript-and-typescript))
+
+```bash
+npm i @desert-ant-labs/tongue
+```
+<!-- card-install:end -->
+
+
+Name the language from three words, across **84 languages**. Built for where detection
+is hardest and most useful on a device: search boxes, chat messages, keyboard input. The
+shipped weights are **2MB** int8 (2,104,940 bytes), there is no tokenizer and no vocabulary
+file, and a detection costs tens of microseconds.
+
+> `"kann ich das haben"` → German · `"안녕하세요"` → Korean · `"привет как дела"` → Russian · `"quanto costa il biglietto"` → Italian
+
+On genuinely ambiguous input it says so rather than guessing: `"la casa"` comes
+back as Italian *or* Spanish, because the phrase is equally both.
+
+**Live demo:** [desert-ant-labs/tongue-demo](https://huggingface.co/spaces/desert-ant-labs/tongue-demo)
+,  runs entirely in your browser; nothing you type leaves the page.
+
+## Files
+
+| File | Format | Size | Contents |
+|---|---|---:|---|
+| `tongue_int8.bin` | Raw int8 + fp32 | 2.01 MiB | The shipped artifact. Byte-identical to what the live demo runs. |
+| `tongue_int4.bin` | Raw int4 + fp32 | 1.01 MiB | Half-size alternative for tight bundles. Costs roughly 0.4pp on short text, see Sizes. |
+| `tongue.onnx` | ONNX (fp32, opset 17) | 8.4MB | Portable graph for onnxruntime / onnxruntime-web. |
+| `tongue_meta.json` | JSON | tiny | Runtime tables: label order, hashing constants, script routing. |
+| `labels.json` | JSON | tiny | The 59 model labels plus the script-decided languages, with English and native names. |
+
+There is no tokenizer file, so nothing has to be shipped or version-matched
+alongside the weights.
+
+## Inputs and outputs
+
+**Input:** a short UTF-8 string, up to 512 characters.
+**Output:** ranked ISO 639-1/639-3 codes with probabilities, plus a reliability
+signal (`confident` / `likely` / `tentative`). Script-decided inputs return a
+single confident answer.
+
+The ONNX graph carries **only the head**: `values` (int64 hashed bucket ids) and
+`offsets` (int64 per-sample starts) in, `logits` out. Normalization, hashing and
+script routing run in the host before the graph; `tongue_meta.json` documents them.
+
+## Coverage
+
+59 languages are learned by the lexical model and a further 25 are decided by
+script alone, across 31 scripts, Latin, Cyrillic, Arabic, Greek and the CJK and
+Indic families among them, for **84 languages** in total. One of the 84,
+Mongolian, is detected only in the traditional Mongolian script; see failure
+mode 3.
+
+## Sizes
+
+Two quantisations of the same weights ship side by side. Pick on bundle budget,
+not on principle.
+
+| | `tongue_int8.bin` | `tongue_int4.bin` |
+|---|---|---|
+| Size | 2.01 MiB | **1.01 MiB** |
+| FLORES 2-word | 0.869 | 0.866 |
+| FLORES 5-word | 0.974 | 0.973 |
+| Held-out single words | 0.759 | 0.752 |
+| Held-out sentences | 0.971 | 0.970 |
+
+The int4 loss lands almost entirely on one- and two-word input; full sentences
+are unaffected within measurement noise. Since short text is what this model is
+for, int8 stays the default and int4 is the option when a megabyte matters more
+than the last half point.
+
+## Failure modes (read before deploying)
+
+Publishing these is part of the product.
+
+**1. One or two words is often genuinely undecidable, and no model size fixes
+it.** A single common word frequently belongs to several languages at once
+(`"sale"` is English, French and Italian; `"la casa"` is equally Italian and
+Spanish). tongue reports a tie or a tentative answer in these cases, and does not
+catch every one: a phrase mixing languages, like `"un garage sale"`, can still
+draw a confident-looking single answer.
+Mitigation: treat low-reliability output as "unknown", not as an answer, and ask
+for more text where the product allows it.
+
+**2. Malay and Indonesian are not reliably separable.** They share vocabulary
+and orthography to the point where short samples carry no distinguishing
+signal. This is a structural limit, not a tuning gap, it is not cheaply
+closable at this size, and every detector we measured struggles with it.
+Mitigation: if you need the distinction, treat `ms`/`id` as one bucket or
+disambiguate from user locale.
+
+**3. Mongolian is detected only in the traditional Mongolian script.**
+Mongolian written in Cyrillic, the dominant modern orthography, is not
+distinguished from the other Cyrillic languages and will usually come back as
+Russian. The language count includes Mongolian because the traditional script
+works; Cyrillic Mongolian does not. Mitigation: do not rely on tongue for
+Cyrillic Mongolian.
+
+**4. Brand names, numbers and code are not language.** `"Samsung Galaxy"`,
+`"v1.2.3"` and `"2024 annual report"` have no correct answer; the model will
+still return its best guess for anything with letters in it. Mitigation: filter
+non-prose input before detection.
+
+**5. Single-word scores are vocabulary recognition, not generalization.** The
+frequent words of a language appear in everyone's training data, so any
+detector's single-word accuracy partly measures memorized vocabulary. Read the
+word-pair and sentence numbers as the generalization signal.
+
+## Measured quality
+
+Every number below is measured on the shipped int8 weights, on three
+public benchmarks, with other detectors run on the identical rows and language
+subsets. Higher is better.
+
+### FLORES-200
+
+Sentences from FLORES-200 truncated to their first 2, 3 and 5 words. Accuracy
+over the 20 languages the three detectors share.
+
+| Detector | Size | 2 words | 3 words | 5 words |
+|---|---|---|---|---|
+| **tongue** | **2MB** | **0.869** | **0.933** | **0.974** |
+| lingua | 293MB | 0.800 | 0.887 | 0.956 |
+| eld | 1MB | 0.780 | 0.856 | 0.912 |
+
+### The lingua test set, the benchmark that library publishes
+
+1,000 single words, word pairs and sentences per language, drawn from the
+same collection lingua trains on. Accuracy over the languages we share.
+
+| Detector | Size | Single words | Word pairs | Sentences |
+|---|---|---|---|---|
+| **tongue** | **2MB** | **0.746** | **0.909** | **0.988** |
+| lingua | 293MB | 0.752 | 0.915 | 0.985 |
+
+### eld, an independent benchmark
+
+Accuracy over the languages tongue supports (53,035 single-word rows,
+53,613 word pairs, 53,141 sentences, 9,066 tweets).
+Apple is the built-in system detector; HeLI-OTS is a 51MB JVM model. lingua 2.2.0 installs as a
+single 293MB compiled extension with its language models embedded.
+
+| Detector | Size | Tweets | Single words | Word pairs | Sentences |
+|---|---|---|---|---|---|
+| **tongue** | **2MB** | **0.992** | **0.759** | **0.887** | **0.971** |
+| lingua | 293MB | 0.984 | 0.756 | 0.894 | 0.950 |
+| HeLI-OTS | 51MB | 0.986 | 0.683 | 0.843 | 0.967 |
+| Apple | system | 0.997 | 0.641 | 0.719 | 0.748 |
+
+## Latency
+
+Measured per single detection in JavaScript on an Apple-silicon laptop: 0.013ms for one word, 0.028ms for a short sentence, 0.10ms at 193 characters (p99 0.24ms). On-device budgets on phone-class hardware will be higher; the design target is under 1ms.
+
+<!-- card-footer:start (generated from manifest.json, edit above this block) -->
+## License
+
+[Desert Ant Labs Source-Available License](https://license.desertant.com/1.0). Free for most
+apps, and a commercial license is required at scale. Full terms are at the link.
+Licensing: <licensing@desertant.com>.
+
+See [`THIRD_PARTY_NOTICES.md`](THIRD_PARTY_NOTICES.md).
+
+## Citation
+
+```bibtex
+@software{tongue_2026,
+  title  = {Tongue: On-device language identification for short text across 84 languages},
+  author = {Desert Ant Labs},
+  year   = {2026},
+  url    = {https://huggingface.co/desert-ant-labs/tongue},
+}
+```
+
+---
+
+© 2026 Desert Ant Labs · <https://desertant.com>
+<!-- card-footer:end -->

THIRD_PARTY_NOTICES.md

@@ -0,0 +1,72 @@
+# Third-party notices — tongue
+
+tongue is trained from scratch and does not derive from any third-party model.
+The training and evaluation corpora below are licensed by their respective
+projects, and those licenses apply to that data. Nothing in the Desert Ant Labs
+Source-Available License overrides them.
+
+Every source used for training is CC0, CC BY, or a permissive software license.
+Share-alike and non-commercial sources are excluded by policy, and the corpus
+build enforces the exclusion and records a provenance manifest of every file
+kept and dropped.
+
+## Training data
+
+### Tatoeba — sentence and link exports
+- **Source:** [tatoeba.org](https://tatoeba.org) — `downloads.tatoeba.org/exports/`
+- **License:** CC BY 2.0 FR
+- **Attribution:** © Tatoeba contributors, licensed under CC BY 2.0 FR
+- **Use in tongue:** the primary corpus. Sentences are windowed into 1–5 and
+  8-word spans, and their distinct tokens supply single-word training rows. The
+  translation-link graph is used to split train/validation by translation family
+  so a sentence and its translations cannot straddle the split.
+
+### Common Voice — sentence collections
+- **Source:** [common-voice/common-voice](https://github.com/common-voice/common-voice) (`server/data`)
+- **License:** CC0 1.0
+- **Use in tongue:** additional sentence-register text, strongest for languages
+  where Tatoeba is thin (Catalan, Basque, Galician, Welsh, Swahili, Belarusian).
+- **Exclusions:** files derived from Wikipedia or Europarl are **not** used, as
+  their upstream sources are share-alike. The fetcher drops any file matching
+  `*wiki*` or `*europarl*` and writes the resulting keep/drop manifest.
+
+### Wikidata Lexemes
+- **Source:** [wikidata.org](https://www.wikidata.org) lexeme dumps
+- **License:** CC0 1.0
+- **Use in tongue:** dictionary-register vocabulary rows (single words).
+
+### Hunspell dictionaries
+- **Source:** [wooorm/dictionaries](https://github.com/wooorm/dictionaries) — English, Dutch, Lithuanian, Russian, Turkish, Persian
+- **License:** per dictionary (MIT / BSD / Apache-2.0 and similar permissive terms; see each dictionary's own license file)
+- **Use in tongue:** additional clean single-word vocabulary for those six languages.
+
+### Universal Dependencies treebanks
+- **Source:** [universaldependencies.org](https://universaldependencies.org)
+- **License:** CC BY 4.0 (attribution required), verified per treebank
+- **Treebanks used:** UD_Spanish-AnCora, UD_Catalan-AnCora, UD_Finnish-FTB,
+  UD_Italian-MarkIT, UD_Portuguese-Porttinari
+- **Use in tongue:** written-register sentences (news, reviews), taken from the
+  `# text =` lines. Share-alike and non-commercial UD treebanks are excluded.
+
+## Evaluation only — never used for training
+
+- **FLORES-200** — NLLB Team et al. — CC BY-SA 4.0 — held out; used for the
+  neutral truncation benchmark.
+- **WiLI-2018** — ODC-BY 1.0 — held out; tail-language coverage check.
+- **eld benchmark** — [nitotm/efficient-language-detector](https://github.com/nitotm/efficient-language-detector) — Apache-2.0 — held out; independent generalization check.
+
+These sets are used to measure the model and are never part of the training
+corpus. Leipzig / Wortschatz corpora are excluded from training in every form
+(news, web, wiki, frequency lists, and mirrors), because another detector's
+published test set is drawn from that collection and training on it would make
+comparisons meaningless.
+
+## Comparison baselines
+
+Numbers reported alongside tongue were produced by running these systems on the
+identical rows and language subsets. They are not redistributed here.
+
+- **lingua** ([pemistahl/lingua-py](https://github.com/pemistahl/lingua-py)) — Apache-2.0
+- **eld** ([nitotm/efficient-language-detector](https://github.com/nitotm/efficient-language-detector)) — Apache-2.0
+- **HeLI-OTS** — [University of Helsinki](https://zenodo.org/record/841984) — used as published
+- **Apple `NLLanguageRecognizer`** — the operating-system detector, measured through a Swift harness

labels.json

@@ -0,0 +1,176 @@
+{
+ "model_labels": {
+  "en": "English",
+  "es": "Spanish",
+  "pt": "Portuguese",
+  "fr": "French",
+  "de": "German",
+  "it": "Italian",
+  "nl": "Dutch",
+  "pl": "Polish",
+  "tr": "Turkish",
+  "sv": "Swedish",
+  "da": "Danish",
+  "no": "Norwegian",
+  "fi": "Finnish",
+  "cs": "Czech",
+  "ro": "Romanian",
+  "hu": "Hungarian",
+  "id": "Indonesian",
+  "vi": "Vietnamese",
+  "ca": "Catalan",
+  "hr": "Croatian",
+  "sk": "Slovak",
+  "lt": "Lithuanian",
+  "et": "Estonian",
+  "af": "Afrikaans",
+  "eu": "Basque",
+  "gl": "Galician",
+  "is": "Icelandic",
+  "tl": "Tagalog",
+  "ms": "Malay",
+  "az": "Azerbaijani",
+  "sl": "Slovenian",
+  "sw": "Swahili",
+  "ga": "Irish",
+  "sq": "Albanian",
+  "cy": "Welsh",
+  "lv": "Latvian",
+  "eo": "Esperanto",
+  "la": "Latin",
+  "ku": "Kurdish",
+  "yo": "Yoruba",
+  "xh": "Xhosa",
+  "st": "Sotho",
+  "ru": "Russian",
+  "uk": "Ukrainian",
+  "bg": "Bulgarian",
+  "sr": "Serbian",
+  "mk": "Macedonian",
+  "be": "Belarusian",
+  "kk": "Kazakh",
+  "ky": "Kyrgyz",
+  "ar": "Arabic",
+  "fa": "Persian",
+  "ug": "Uyghur",
+  "ur": "Urdu",
+  "hi": "Hindi",
+  "mr": "Marathi",
+  "ne": "Nepali",
+  "bn": "Bengali",
+  "as": "Assamese"
+ },
+ "script_decided": {
+  "am": "Amharic",
+  "bo": "Tibetan",
+  "chr": "Cherokee",
+  "dv": "Dhivehi",
+  "el": "Greek",
+  "gu": "Gujarati",
+  "he": "Hebrew",
+  "hy": "Armenian",
+  "ja": "Japanese",
+  "ka": "Georgian",
+  "km": "Khmer",
+  "kn": "Kannada",
+  "ko": "Korean",
+  "lo": "Lao",
+  "ml": "Malayalam",
+  "mn": "Mongolian",
+  "my": "Burmese",
+  "or": "Odia",
+  "pa": "Punjabi",
+  "si": "Sinhala",
+  "syr": "Syriac",
+  "ta": "Tamil",
+  "te": "Telugu",
+  "th": "Thai",
+  "zh": "Chinese"
+ },
+ "native_names": {
+  "en": "English",
+  "es": "Español",
+  "pt": "Português",
+  "fr": "Français",
+  "de": "Deutsch",
+  "it": "Italiano",
+  "nl": "Nederlands",
+  "pl": "Polski",
+  "tr": "Türkçe",
+  "sv": "Svenska",
+  "da": "Dansk",
+  "no": "Norsk",
+  "fi": "Suomi",
+  "cs": "Čeština",
+  "ro": "Română",
+  "hu": "Magyar",
+  "id": "Indonesia",
+  "vi": "Tiếng Việt",
+  "ca": "Català",
+  "hr": "Hrvatski",
+  "sk": "Slovenčina",
+  "lt": "Lietuvių",
+  "et": "Eesti",
+  "af": "Afrikaans",
+  "eu": "Euskara",
+  "gl": "Galego",
+  "is": "Íslenska",
+  "tl": "Tagalog",
+  "ms": "Melayu",
+  "az": "Azərbaycan",
+  "sl": "Slovenščina",
+  "sw": "Kiswahili",
+  "ga": "Gaeilge",
+  "sq": "Shqip",
+  "cy": "Cymraeg",
+  "lv": "Latviešu",
+  "eo": "Esperanto",
+  "la": "Latina",
+  "ku": "Kurdî",
+  "yo": "Yorùbá",
+  "xh": "isiXhosa",
+  "st": "Sesotho",
+  "ru": "Русский",
+  "uk": "Українська",
+  "bg": "Български",
+  "sr": "Српски",
+  "mk": "Македонски",
+  "be": "Беларуская",
+  "kk": "Қазақ",
+  "ky": "Кыргыз",
+  "ar": "العربية",
+  "fa": "فارسی",
+  "ug": "ئۇيغۇر",
+  "ur": "اردو",
+  "hi": "हिन्दी",
+  "mr": "मराठी",
+  "ne": "नेपाली",
+  "bn": "বাংলা",
+  "as": "অসমীয়া",
+  "am": "አማርኛ",
+  "bo": "བོད",
+  "chr": "ᏣᎳᎩ",
+  "dv": "ދިވެހި",
+  "el": "Ελληνικά",
+  "gu": "ગુજરાતી",
+  "he": "עברית",
+  "hy": "Հայերեն",
+  "ja": "日本語",
+  "ka": "ქართული",
+  "km": "ខ្មែរ",
+  "kn": "ಕನ್ನಡ",
+  "ko": "한국어",
+  "lo": "ລາວ",
+  "ml": "മലയാളം",
+  "mn": "Монгол",
+  "my": "မြန်မာ",
+  "or": "ଓଡ଼ିଆ",
+  "pa": "ਪੰਜਾਬੀ",
+  "si": "සිංහල",
+  "syr": "ܣܘܪܝܝܐ",
+  "ta": "தமிழ்",
+  "te": "తెలుగు",
+  "th": "ไทย",
+  "zh": "中文"
+ }
+}

tongue.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:22e3d33efa6b0cf4c75401e3f5becb243ef5a4256d48d8b399e56eb62471359e
+size 8399496

tongue_int4.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a8067e66bca3b254f41265d020d87f98127469d6f29101fb627f7c40e5414c4d
+size 1056492

tongue_int8.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2cea7b6a92be043c58596d0b6f482fbc27fb901e5697d77159247d3ec6bdf767
+size 2104940

tongue_meta.json

@@ -0,0 +1,197 @@
+{
+ "labels": [
+  "en",
+  "es",
+  "pt",
+  "fr",
+  "de",
+  "it",
+  "nl",
+  "pl",
+  "tr",
+  "sv",
+  "da",
+  "no",
+  "fi",
+  "cs",
+  "ro",
+  "hu",
+  "id",
+  "vi",
+  "ca",
+  "hr",
+  "sk",
+  "lt",
+  "et",
+  "af",
+  "eu",
+  "gl",
+  "is",
+  "tl",
+  "ms",
+  "az",
+  "sl",
+  "sw",
+  "ga",
+  "sq",
+  "cy",
+  "lv",
+  "eo",
+  "la",
+  "ku",
+  "yo",
+  "xh",
+  "st",
+  "ru",
+  "uk",
+  "bg",
+  "sr",
+  "mk",
+  "be",
+  "kk",
+  "ky",
+  "ar",
+  "fa",
+  "ug",
+  "ur",
+  "hi",
+  "mr",
+  "ne",
+  "bn",
+  "as"
+ ],
+ "num_buckets": 65536,
+ "dim": 32,
+ "ngram_orders": [
+  1,
+  2,
+  3,
+  4,
+  5
+ ],
+ "whole_tokens": false,
+ "embed_scale": 0.06248359229620986,
+ "layout": {
+  "emb_int8": [
+   65536,
+   32
+  ],
+  "lin_w_f32": [
+   59,
+   32
+  ],
+  "lin_b_f32": [
+   59
+  ]
+ },
+ "latin_labels": [
+  "en",
+  "es",
+  "pt",
+  "fr",
+  "de",
+  "it",
+  "nl",
+  "pl",
+  "tr",
+  "sv",
+  "da",
+  "no",
+  "fi",
+  "cs",
+  "ro",
+  "hu",
+  "id",
+  "vi",
+  "ca",
+  "hr",
+  "sk",
+  "lt",
+  "et",
+  "af",
+  "eu",
+  "gl",
+  "is",
+  "tl",
+  "ms",
+  "az",
+  "sl",
+  "sw",
+  "ga",
+  "sq",
+  "cy",
+  "lv",
+  "eo",
+  "la",
+  "ku",
+  "yo",
+  "xh",
+  "st"
+ ],
+ "script_decisive": {
+  "Hangul": "ko",
+  "Hiragana": "ja",
+  "Katakana": "ja",
+  "Thai": "th",
+  "Lao": "lo",
+  "Khmer": "km",
+  "Myanmar": "my",
+  "Greek": "el",
+  "Hebrew": "he",
+  "Armenian": "hy",
+  "Georgian": "ka",
+  "Ethiopic": "am",
+  "Cherokee": "chr",
+  "Tamil": "ta",
+  "Telugu": "te",
+  "Kannada": "kn",
+  "Malayalam": "ml",
+  "Sinhala": "si",
+  "Gujarati": "gu",
+  "Gurmukhi": "pa",
+  "Oriya": "or",
+  "Thaana": "dv",
+  "Tibetan": "bo",
+  "Mongolian": "mn",
+  "Syriac": "syr",
+  "Han": "zh"
+ },
+ "script_narrowing": {
+  "Cyrillic": [
+   "ru",
+   "uk",
+   "bg",
+   "sr",
+   "mk",
+   "be",
+   "kk",
+   "ky"
+  ],
+  "Arabic": [
+   "ar",
+   "fa",
+   "ur",
+   "ug"
+  ],
+  "Devanagari": [
+   "hi",
+   "mr",
+   "ne"
+  ],
+  "Bengali": [
+   "bn",
+   "as"
+  ]
+ },
+ "normalization": "NFC, lowercase, strip URLs/emails/mentions/digits, collapse whitespace, cap 512 chars",
+ "hashing": "FNV-1a 32-bit over Unicode scalars of '^token$', n-gram orders as listed, modulo num_buckets",
+ "onnx": {
+  "file": "tongue.onnx",
+  "inputs": {
+   "values": "int64[num_ngrams] — fnv1a(ngram) % num_buckets, flat over the batch",
+   "offsets": "int64[batch] — start index of each sample within values"
+  },
+  "output": "logits float32[batch, num_labels] in `labels` order",
+  "note": "head only; normalization, hashing and script routing run in the host"
+ }
+}